AI-Powered Property Tax Query Resolution Pilot Model for Smart City Using DeepQuery

1. Background & Objective

Bilaspur, one of the 100 Smart Cities selected under the National Smart Cities Mission, aimed to improve citizen-centric digital services—starting with property tax. Historically, property tax queries were managed via static websites and in-person counters, both of which failed to provide real-time, comprehensible, multilingual support to residents.

To address this, Bilaspur Smart City Ltd. partnered with Presear Softwares Pvt. Ltd. to deploy DeepQuery, a multilingual AI-powered assistant trained specifically for municipal property tax services.

Objective:

• Enable 24×7 query resolution for property tax via WhatsApp and web.

• Simplify legal provisions and policy explanations using natural language.

• Increase tax compliance and citizen satisfaction using AI.

---

2. Dataset Preparation & Knowledgebase Design

The bot’s core intelligence was developed from statutory documents and real citizen pain points.

Key Sources:

• Chhattisgarh Municipal Corporation Act – Part IV Chapter XI: Taxation

• Property tax demand notices, rebate policies, penalty circulars

• Citizen RTI responses and past FAQs from Bilaspur Municipal Corporation

• Internal municipal SOPs and mutation workflows

Knowledgebase Highlights:

• 300+ bilingual Q&A pairs created manually

• Covered intents like tax calculation, rebates, penalty, payment deadlines, receipts, mutation status

• Structured into contextual categories for retrieval

Example Intent Coverage:

• "How much tax do I owe this year?"

• "क्या छूट मार्च के बाद मिलती है?"

• "I paid online but didn’t get a receipt."

---

3. Language Support & NLP Tuning

Given Bilaspur’s linguistic diversity, DeepQuery was built with multilingual capability at its core.

Approach:

• All content was translated into Hindi with contextual integrity.

• Hinglish and misspelled inputs were normalized using phonetic matching.

• Regional dialect queries were supported via synonym mapping (e.g., “bhugtan”, “jama”, “kar”).

Enhancements:

• Used custom embeddings for semantic similarity in both Hindi and English

• Integrated fallback keyword detection to handle out-of-scope queries gracefully

• Voice-to-text (STT) and text-to-speech (TTS) modules were added for accessibility

---

4. Model Architecture & Training

DeepQuery used a hybrid Retrieval-Augmented Generation (RAG) approach, tuned for government document comprehension.

Architecture:

• Embedding-based semantic search (using in-house vector database)

• Transformer-based response generation tuned on government corpus

• Custom logic for contextual grounding (e.g., ward-specific rules, due dates)

Training Loop:

• Initial supervised Q&A-based fine-tuning

• Weekly retraining with new queries from real usage

• Feedback from BMC officials and citizens used to refine answer sets

---

5. Integration & Deployment

Platforms:

• Embedded Web Widget on Bilaspur Smart City official portal

---

6. Results & Impact

Within 60 days of deployment:

• 25,000+ citizen queries handled

• < 5 seconds average response time

• 92% satisfaction score

• Significant reduction in load on municipal helplines and counters

Citizen Experience:

• Queries answered in natural language

• Voice support enabled for illiterate and elderly users

• Proactive reminders about rebate expiry via WhatsApp

---

7. Challenges & Learnings

---

8. Alignment with Cities Innovation Exchange (CiX)

DeepQuery directly supports CiX goals of scalable, AI-powered urban solutions.

• Demonstrates scalable use of LLMs in urban governance

• Promotes inclusion via multilingual, multimodal access

• Reduces human dependency for high-volume citizen services

• Builds a replicable template for other urban local bodies in India

---

9. Way Forward

• Expansion into other domains: water tax, building approvals, trade licenses

• Automated multilingual reminders for tax deadlines

• Integration with Digital Property Ledger for real-time mutation updates

• Real-time grievance lodging and tracking within the same AI assistant